Process
Agree the bar. Then clear it.
Every project follows one line: quality, measured against a bar we set with you at the start. Five stages take it from a first prototype to a system that keeps improving in production, and you can see where it stands at every step.
Five stages, one curve
Scroll through a project. The line is an example run: quality at each checkpoint, against the bar we agree at the start. Each stage ends with something you can hold, and a call that is yours to make.
- 01
We sit with the people who do the work, map the workflow and the data behind it, and agree what good looks like in terms you can check. That becomes the bar everything after is measured against.
- You get
- A one-page map of the workflow and a written quality bar.
- You decide
- Which workflow to start with, and where the bar sits.
Workflow mapExample
- A request comes in
- Find the facts
- Draft the answer
- A person signs off
Data: past requests, policy documents, the order system
- 02
Within weeks you see a working demo on your real data, not a mock-up. It is rough on purpose: the point is to learn quickly where the AI is strong and where it struggles.
- You get
- A working demo you can try with your own cases.
- You decide
- Whether it is worth measuring properly.
Working demoExample
Can I still return the jacket I ordered last month?
Yes. Your order is still inside the return window, so I have started the return and emailed you a label.
returns-policy.pdfon your data
- 03
We measure accuracy, cost, and speed against the bar we agreed, on test cases drawn from your real work. You read the same report we do, failures included.
- You get
- An eval report: every test case, pass or fail, and the score against the bar.
- You decide
- Whether it is good enough to build, or what to fix first.
Eval reportExample
passfailreal test cases
score0.77 - 04
Production-grade integration with your systems, with security and monitoring built in from the start. The evals come along, so every release is checked against the bar before it ships.
- You get
- A production system connected to your tools, with monitoring on.
- You decide
- Who can use it, what it can touch, and when it goes live.
Production checklistExample
- Connected to your systems
- Access control and audit log
- Evals run on every release
- Monitoring and alerts on
Ready to go live
- 05
We watch it in production and keep making it better: we catch drift, fix the cases it gets wrong, and switch models when a better one arrives. Each improvement is measured before it ships.
- You get
- A live monitor, and a note each time an improvement ships.
- You decide
- Which improvements matter most to your team.
Live monitorExample
quality in productionAbove the bar
- Improvement shippedBetter answers on returns
- Improvement shippedLower cost per task
- Improvement shippedMoved to a newer model
Workflow map01 · Discover
- A request comes in
- Find the facts
- Draft the answer
- A person signs off
Data: past requests, policy documents, the order system
Working demo02 · Prototype
Can I still return the jacket I ordered last month?
Yes. Your order is still inside the return window, so I have started the return and emailed you a label.
returns-policy.pdfon your data
Eval report03 · Evaluate
passfailreal test cases
Production checklist04 · Build
- Connected to your systems
- Access control and audit log
- Evals run on every release
- Monitoring and alerts on
Ready to go live
Live monitor05 · Run & improve
quality in productionAbove the bar
- Improvement shippedBetter answers on returns
- Improvement shippedLower cost per task
- Improvement shippedMoved to a newer model
How you stay in the loop
You see the work as it happens, and anything that needs your judgement is marked as yours to decide.
- Needs your decision
- Done
Project channelWhat a week looks like · example
- Done
CraftsAIWeekly demo
A short recorded walkthrough of this week's version, including the cases it still gets wrong.
- Done
CraftsAIEval report
The new version clears the bar on the latest test cases. The remaining failures are listed with examples.
- Needs your call
CraftsAIDecision log
Should the agent approve small refunds on its own, or always ask a person first? The options and trade-offs are written up.
- Done
CraftsAIDeploy note
The new version is on staging for your team to try. The previous one stays ready if anything looks off.
What we won't do
Demos that can't ship.
Every prototype runs on your real data, so what you see is what we build on.
Launches without evals you can read.
Nothing goes live until it clears the bar, and you get the report.
Black boxes.
You own the code, the prompts, the evals, and the data.
Start with Discover.
Tell us the workflow. We'll map it with you, agree the bar, and give you an honest view of whether AI can clear it.